Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Remove FSDP wrapping from sub-models. #34452

Open
wants to merge 6 commits into
base: main
Choose a base branch
from
+33 −3
Open

Remove FSDP wrapping from sub-models. #34452

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
ab9934e
Remove FSDP wrapping from sub-models.
eljandoubi Oct 27, 2024
d518f76
solve conflict trainer.py
eljandoubi Oct 28, 2024
e76dfd0
make fixup
eljandoubi Oct 29, 2024
a6dcb92
add unit test for fsdp_auto_wrap_policy when using auto_find_batch_size
eljandoubi Oct 30, 2024
7ebb2c6
put back extract_model_from_parallel
eljandoubi Nov 4, 2024
0df20d6
use transformers unwrap_model
eljandoubi Nov 4, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 6 additions & 3 deletions src/transformers/trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -66,7 +66,7 @@
from .integrations.deepspeed import deepspeed_init, deepspeed_load_checkpoint, is_deepspeed_available
from .integrations.tpu import tpu_spmd_dataloader
from .modelcard import TrainingSummary
from .modeling_utils import PreTrainedModel, load_sharded_checkpoint
from .modeling_utils import PreTrainedModel, load_sharded_checkpoint, unwrap_model
from .models.auto.modeling_auto import (
MODEL_FOR_CAUSAL_LM_MAPPING_NAMES,
MODEL_MAPPING_NAMES,
Expand Down Expand Up @@ -2277,8 +2277,11 @@ def _inner_training_loop(
# FSDP-XLA, SageMaker MP/DP, DataParallel, IPEX
use_accelerator_prepare = True if model is self.model else False

# configure fsdp plugin for qlora if any
if use_accelerator_prepare:
if use_accelerator_prepare and self.is_fsdp_enabled:
# In case of auto_find_batch_size=True
# Remove FSDP wrapping from sub-models.
self.model = unwrap_model(self.model, recursive=True)
# configure fsdp plugin for qlora if any
self._fsdp_qlora_plugin_updates()

if delay_optimizer_creation:
Expand Down
27 changes: 27 additions & 0 deletions tests/trainer/test_trainer_fsdp.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -117,6 +117,33 @@ def test_trainer(self):
execute_subprocess_async(cmd, env=self.get_env())
# successful return here == success - any errors would have caused an error in the sub-call

class TestFSDPTrainerWrap(TestCasePlus):
@require_accelerate
@require_torch_multi_gpu
@require_fsdp
def test_trainer(self):
output_dir = self.get_auto_remove_tmp_dir()
cmd = [
"accelerate",
"launch",
"--use_fsdp",
"--main_process_port",
f"{get_torch_dist_unique_port()}",
"--num_processes",
f"{torch.cuda.device_count()}",
"--fsdp_transformer_layer_cls_to_wrap",
"GPT2Block",
f"{self.test_file_dir}/test_trainer_fsdp.py",
"--output_dir",
f"{output_dir}",
"--report_to",
"none",
"--auto_find_batch_size",
"True",
]
execute_subprocess_async(cmd, env=self.get_env())
# successful return here == success - any errors would have caused an error in the sub-call


if __name__ == "__main__":
parser = HfArgumentParser((Seq2SeqTrainingArguments,))
Expand Down